Today, three connected pages from the Hugging Face Agents Course:
- Understanding AI Agents through the Thought-Action-Observation Cycle
- Thought: Internal Reasoning and the ReAct Approach
- Actions, Enabling the Agent to Engage with Its Environment
This finally ties Days 3 and 4 together into one loop. An agent runs Thought → Action → Observation, repeatedly: it reasons about what to do next (Thought), calls a tool (Action), reads the result back into context (Observation), and repeats until done. ReAct is the prompting pattern that makes this work — interleaving reasoning with tool calls instead of doing all the thinking up front, so the agent can react to what it actually learns along the way (as opposed to Chain-of-Thought, which just reasons step by step with no environment interaction).
On the Action side, the concrete mechanism is the "stop and parse" approach: the LLM outputs its intended action in a structured format, generation is stopped right there, and an external parser reads the structure and actually executes the tool — the model itself never runs anything. Agents built this way fall into a spectrum: JSON agents (simple, structured, easy to validate) vs. Code agents (write executable code, more expressive with loops/conditionals, but need sandboxing since they can do more damage if wrong).
Put together: the loop is the architecture, ReAct is the reasoning strategy inside it, and Action format (JSON vs. code) is just an implementation choice for the "Action" step.